Founding Engineer

Trade Me Role Description

Millions of Kiwi workers are stuck. Not because there are no jobs - there are. The system for finding them is broken. Meanwhile local small businesses struggle to hire, choked by slow, manual processes. We're fixing both sides.

We're a small, fast startup team inside Trade Me Jobs: NZ's largest jobs platform, 4 million members, data no competitor can replicate. We're building an AI service that connects blue-collar workers - the heroes of our economy - with real opportunities before they think to look.

Why this role is different

• A founding team of two: a product founder and an engineering founder moving at extreme speed. You're our first permanent hire.
• Trade Me security and full benefits, with the autonomy and ownership of an early-stage startup.
• Your fingerprints on everything we build, from day one.

What you'll work on Core product engineering and rapid market experiments - you'll be close to all of it.

• For workers: how they discover jobs, put themselves forward, improve their chances and stay engaged. Surfaces that feel immediate and natural - custom SMS, personalised temporary pages, natural language instead of forms.
  - For employers (builders, plumbers, electricians, transport fleets with hard-to-fill roles): the engines behind rapid, high-integrity experiments - listing creation, change detection, candidate parsing, matching, notification loops.

The tech under the hood

• Python and Postgres do the heavy lifting. FastAPI and Pydantic everywhere, Pydantic AI driving the LLM parts, Gemini on Vertex behind it.
• Semantic search (pgvector), a knowledge graph, and plenty of testing and guardrailing to keep LLM behaviour honest. You needn't be a graph or search specialist, but you must be comfortable when a model returns something plausible and wrong.
• Svelte and Skeleton UI on the front end — you should care how the thing feels to use.
• An AI agent runs against live production data daily, built on Pi (a minimal open-source coding agent) with our own tooling on top. Operations prove themselves there first, then graduate into the backend as background workers. It's yours to shape.
• Fluency with a harness (Claude Code, Codex, whatever) as your normal way of building, with opinions on doing it well.

The ground shifts every few months and none of us has this figured out. Technical creativity counts more here than depth in any one of the above.

Who you are
You've built products before, not just features.

• Product instinct. You think about why as much as how, you're curious about customers' daily lives, and you move on ambiguity rather than wait for a perfect spec. You've been on the phone to customers.
• True full-stack. Data heavy: highly active across our typed Python backend (FastAPI, Pydantic) and into the frontend. Data wrangling helps; ML or NLP specialism isn't required.
• AI native. Using AI to write code and test ideas is your baseline, not a bonus. You've built with LLMs: prompt design, structured output parsing, context optimisation.
  - Strict quality. TDD is non-negotiable. We lean on pytest unit and integration tests to ship at pace.

Nice to have: you've run an LLM system in production and dealt with it being confidently wrong, or talked a small business into trying something unproven.

What we don't want
A perfect CV describing maintenance of massive legacy systems. We want someone who builds new things, gets comfortable when they break, learns fast from real users, and ships again.

A window into the role
We didn't write this section. We pointed Claude Code at three and a half months of our session logs, asked what we do well and what would make it more effective, then sent more agents in to disprove them. What survived:

What it reckons we're good at

• They check whether the thing already exists before building it - four short questions once turned thirteen proposed tables into six.
• Nothing generated reaches a customer without provenance.
• They catch me being confidently wrong, seventeen times. I need more of this, not less.

What it reckons we need

• A scoreboard for the fuzzy stuff. One similarity cutoff got set by eyeballing a score at 0.73 and deciding it felt high.
• Someone who runs it before saying it works. Forty tasks landed on main with "no deployment or integration verification", including a sync script silently dropping 96% of records.
• Suspicion of the deploy layer. Staging served a two-month-old build for weeks while every pipeline reported green.

That's the job: product-focused, while improving an evolving new way of working.

The honest pitch
You'll join a founding team with deep experience in search, data, ML and small business. We move fast, we learn in public with the customers we exist for, and we're building something that matters for millions of Kiwi.

You'll have the security of a Trade Me contract - full benefits, leave entitlements, competitive pay - but you'll operate like a startup hire. Direct access to decision-making, zero layers, and no two days the same.